ACF-SAP: A machine learning framework for predicting obstructive sleep apnea severity using anthropometric and clinical features

Abstract

Objective:

This study aims to develop and evaluate ACF-SAP, a machine learning (ML) framework for predicting obstructive sleep apnea (OSA) severity using non-invasive, routinely collected clinical features.

Methods:

The proposed approach leverages common anthropometric and clinical variables, including sex, body mass index (BMI), height, weight, neck circumference, and nocturia. The methodology integrates machine-learning–based feature selection to identify the most informative predictors, followed by unsupervised clustering to generate data-driven sleep severity labels. These labeled data are then used to train and evaluate the ACF-SAP framework.

Results:

ACF-SAP, implemented with ensemble classifiers, achieved a classification accuracy of 0.84, with strong F1-scores and balanced sensitivity across OSA severity levels.

Conclusions:

The ACF-SAP model supports early identification of patients at high risk for OSA and may serve as a first-line screening tool to prioritize referrals for polysomnography (PSG).

Significance:

This work presents a scalable, low-cost screening framework that can improve triage efficiency and facilitate timely diagnosis, particularly in resource-constrained healthcare environments.